Glossary
Enterprise AI Gateway & LLM DLP Glossary
Plain-English definitions for the infrastructure, security and cost-control terms behind ContextMiner: Enterprise AI Gateway, LLM Proxy, LLM Data Loss Prevention (DLP), Prompt Compression & Caching and API Gateway Guardrails.
- Enterprise AI Gateway
- A single control point between your applications and upstream LLM providers that enforces cost, security and compliance policies on every request. ContextMiner is a drop-in, OpenAI-compatible Enterprise AI Gateway.
- LLM Proxy
- A reverse proxy that receives OpenAI-formatted requests, optimizes and protects the payload, then forwards it to the model provider (OpenAI, Anthropic, Google Gemini, DeepSeek, Groq, Mistral AI, OpenRouter or self-hosted engines) and returns the response in the original format.
- LLM Data Loss Prevention (DLP)
- Controls that stop sensitive data from leaving your environment inside prompts. ContextMiner detects personal identifiers and secrets locally and replaces them with reversible tokens before the request goes upstream.
- PII Masking
- Replacing personally identifiable information (CPF, CNPJ, SSN, emails, card numbers, API keys and more) with placeholders or vault tokens. The original values are restored in memory when the model responds.
- Zero-Egress Architecture
- A processing model in which prompts and responses are transformed in memory and never persisted. Only numeric telemetry such as tokens saved and request counts is recorded.
- Prompt Compression
- Reducing the number of tokens sent to a model without changing the task. ContextMiner combines the Dynamic Compression Engine, Structural Logic Analyzer, Deterministic Payload Minifier and Algorithmic Relevance Filter to cut 15% to 80%+ of input tokens depending on plan and mode.
- Prompt Caching
- Serving repeated requests from an in-memory response cache to avoid paying for identical upstream calls. ContextMiner keeps a 1-hour cache on Pro and a 4-hour cache on Enterprise.
- API Gateway Guardrails
- Policy enforcement at the gateway layer: request quotas, multimodal media blocking, SSRF and loop protection, encrypted provider keys (BYOK) and ReDoS-safe custom pattern rules.
- Semantic Skeleton Extractor
- The Web Noise Filter that strips scripts, styles, navigation chrome and layout wrappers from HTML, keeping meaningful text, links and tables before the content reaches the model.
- Algorithmic Relevance Filter
- Also called the Semantic Scaling Module. It splits long text into overlapping chunks, scores them against the user intent and keeps the most relevant ones in their original order, protected by a retention floor.
- Structural Logic Analyzer
- Also called the Safe Syntax Parser. It removes comments and formatting overhead from code while preserving function signatures, imports, docstrings and tool schemas, so coding assistants keep working.
- Deterministic Payload Minifier
- The Structural Data Optimizer that minifies embedded JSON, CSV/TSV data and historical tool-call arguments while guaranteeing the output stays valid.
- Retention Floor
- A guaranteed minimum share of the original tokens that is kept after optimization: 45% in IDE & Dev Mode and 20% in General & RAG Mode, preventing context collapse.
- BYOK (Bring Your Own Key)
- You connect your own provider API keys. ContextMiner encrypts them at rest with AES-256-GCM and decrypts them only when calling the upstream provider.
